On Constructing Vessel Dataset Structure Using GAN-based Data Augmentation
Ah Reum Oh, Jiwon Lee, Sung-Won Moon, Jung-Soo Lee, Do-Won Nam, Wonyoung Yoo · 2021 International Conference on Information and Communication Technology Convergence (ICTC) · 2021
Conventional methods using classical image processing techniques is to limitedly augment basic data for extension of the dataset volume in deep learning network system. A new proposed approach using synthetic data by 3D virtual model and Generative Adversarial Networks (GAN) is able to resolve the lack of dataset adequately, and the performance of the dataset structure can be verified by a classification network model. The single and combined data groups with various types of images were constructed for the accuracy comparison of classification system, and it indicated that the proposal has an appropriate profit for improvement of the system. The composed dataset using data augmentation methods can be applied on both academic and industrial field which have little actual data for deep leaning network systems. Further work will aim to improve the quality of the data from GAN and find the relevant quantity of dataset according to data type.